Self-rectifying memristor boosts positive current to push neural network benchmarks

Memristors execute computations within memory devices without requiring processors and are one potential avenue for scaling up neuromorphic hardware. One implementation, called the self-rectifying memristor (SRM), shows promise, but many approaches have involved a trade off between rectification ratios — a measure of scalability — and linearity — a measure of how precisely synaptic weights can be updated.

Zhang et al. developed a new SRM architecture that reduces cross-talk and increases capacity for precise synaptic weight updates in neural networks, called Pt/HfO₂/WO3-x/TiN. In experiments, the group found that traps in the WO_3-x resistive layer work synergistically with insulation from the HfO₂ layer to suppress negative current while promoting positive current, ultimately boosting performance.

“The ability of the proposed SRMs to maintain high rectification addresses a significant challenge in their development for in-memory computing hardware while simultaneously achieving great potential for ultra-large-scale integration and linearity,” said author Yishu Zhang. “Additionally, our work provides a comprehensive understanding of the underlying mechanisms governing the rectification behavior in SRMs, which is vital for future design optimization and scalability assessments.”

According to the simulation, the passive crossbar array with the proposed SRMs can scale up to over 21 gigabits without misreading. Benefiting from a near-perfect linearity, the synaptic array can realize 98.1% classification accuracy. By comparison, a recent SRM showed a similar linearity but could only scale to 144 megabits.

The group next looks to scale up their memristor arrays as well as integrating their structure into more complex neuromorphic designs.

Source: “Self-rectifying memristors with high rectification ratio and dynamic linearity for in-memory computing,” by Guobin Zhang, Zijian Wang, Xuemeng Fan, Zhen Wang, Pengtao Li, Qi Luo, Dawei Gao, Qing Wan, and Yishu Zhang, Applied Physics Letters (2024). The article can be accessed at https://doi.org/10.1063/5.0225833 .